Ok, then it doesn't suggest a Blender issue. I wasn't sure anymore about this, and its better to double check trivial things, no matter the problem at hand. Long threads with unreproducible errors can get hard to followjgrover110 wrote: ↑Mon Apr 13, 2020 8:41 pm From my testing, I have the same issue with the standalone build as the blender addon. If you are unable to reproduce it, then this must be an issue with my system. The best I can do it describe the issue I am having and investigate what I believe to be the problem.
Do you mean the "importance" setting in the performance tab? That would not be multiple importance sampling. Or is there another importance setting?jgrover110 wrote: ↑Mon Apr 13, 2020 8:41 pm From what I can see, the problem seems to be the sampling of the HDRi in direct lighting, as the brightness of the shadows appears to be correct in the reflection of the metal cube. This was confirmed to me when I tried it with importance of 0.
What this did was effectively disabled importance sampling of the HDRi and was sampling it purely at random. In OpenCL the direct lighting was almost completely black apart from the fireflies where it had sampled the sun. The reflections of the floor however, were a lot brighter, meaning that the sky was being sampled properly. In CPU, the floor was the same brightness as the reflections of the floor, which means that the sky was being sampled properly even in direct light.
Anyway, have you tried the random sampler instead of sobol? Might be worth a try.
Other than that, I tried something to check if it only samples one hemisphere of the HDRI for direct diffuse bounces. As the rest of the floor is sampled well by the sunlight, it can't be general for a type of bounce. I tested it by putting a half-box of black material aroudn the scene. however, that still couldn't quite replicate the look you get... but who knows what regions exactly would have to be left out.